The Impact of Oversampling Techniques for Severe Class Imbalance in Credit Card Fraud Detection

Asaddulloh B.P.; Nafea A.A.; Aminuddin A.; Fakhreldin M.; Enriquez H.R.; Ernawan F. · Proceeding - 2025 IEEE 9th International Conference on Software Engineering and Computer Systems: Advancements in Next-Generation Intelligent Solution, ICSECS 2025 (2025) · DOI: 10.1109/icsecs65227.2025.11279210

Resumen

This paper addresses the critical challenge of detecting credit card fraud in a highly imbalanced dataset of European credit card transactions from September 2013, containing only 0.172% of fraudulent cases. To mitigate the severe class imbalance, three oversampling techniques Synthetic Minority Oversampling Technique (SMOTE), Adaptive Synthetic Sampling Approach (ADASYN), and Random Oversampling (ROS) were employed, successfully balancing the minority class. Four classification algorithms Random Forest, K-Nearest Neighbors (KNN), Gradient Boosting, and AdaBoost were then trained and evaluated. The results demonstrate that models trained on the original, unbalanced dataset exhibited high accuracy but poor recall for fraudulent transactions (KNN with 0.0333 recall). In contrast, the application of oversampling techniques drastically improved performance. Notably, Random Forest combined with Random Oversampling achieved near-perfect scores across all metrics, including a 1.0 ROC AUC and 1.0 recall on the testing set, indicating its exceptional ability to identify fraudulent activities. SMOTE also yielded substantial improvements, particularly for Random Forest and KNN, showcasing high recall (0.99998) and F1-score (0.99989). This research highlights the indispensable role of data balancing in developing robust and effective credit card fraud detection systems.

Instituciones: Universidad Nacional Micaela Bastidas de Apurímac